Data Semantics Revisited: Databases and the Semantic Web

1. Data Semantics Revisited:Databases and the Semantic Web John Mylopoulos University of Toronto Seminar Series on the Semantic Web Univ. of Rome “La Sapienza”, Dept. of Informatics, and LEKS, IASI-CNR Rome, December 9, 2003

2. Next Generation Database Systems Won't Work Without Semantics! Data Semantics Can't Fail This Time! Panel Discussion ACM SIGMOD’98, Seattle, June 4, 1998 Panel Discussion CAiSE’98, Pisa June 11, 1998 …1998…

3. The Panelists Panel: Philip Bernstein (Microsoft) Umesh Dayal (HP Laboratories) Sham Navathe (Georgia Tech) Marek Rusinkiewicz (MCC) Panel chair: John Mylopoulos (Univ. of Toronto) Panel: Michael Brodie (GTE Labs) Stefano Ceri (Politecnico di Milano) Arne Solvberg (Univ. of Trondheim) Panel chair: John Mylopoulos (Univ. of Toronto)

4. “…The three most important problems in Databases used to be Performance, Performance and Performance; in the future, the three most important problems will be Semantics, Semantics and Semantics…” (paraphrase) Stefano Ceri June 11, 1998

5. This Talk • Data Semantics: The problem and its history • The Semantic Web: The vision and the challenges • Towards a new theory of data semantics

6. Data Semantics • Establish and maintain the correspondence between a data source, hereafter a model, and its intended subject matter. • The model may be • A database storing data about employees in a company; • A database schema describing parts, projects and suppliers; • A website presenting information about a university; • A plain text file describing the battle of Waterloo.

7. Machine State vs World Semantics queries/updates Data source Subject matter

8. Semantic Data Models • Data models that attempt to capture “more world knowledge” [Codd79] than their logical counterparts. • Make ontological assumptions about the subject matter, offer primitives accordingly. • For example, the Entity-Relationship model “assumes that the world consists of entities and relationships” [Chen76]. • The Relational Model makes no ontological assumptions [Codd70].

9. History • First semantic data models were proposed in 1974: • Jean-Robert Abrial; • Giampio Bracchi, Paolo Paolini and Giuseppe Pelagatti; • Jean-Luc Hainaut and Alain Pirotte; • Hans Schmid and Richard Swenson; Then in 1975 • Peter Chen; • Nick Roussopoulos and John Mylopoulos; • ….many others…

10. Where do Semantic Data Models Fit? • Several possibilities, actually: • They are part of the technology --> Semantic DBMSs; • They are used during design; • They are part of the user interface to the database. • Option 2 prevailed. Semantics were to be dealt with during design-time, rather than run-time, for performance reasons. • But how does one use a database where semantics have been factored out? Rely on a few users and application programs to know these semantics

11. …There is a down side to this data management practice: Legacy data…

12. What did the Panel Experts See? • Factoring out the semantics of data won’t work in dynamically changing, distributed, open environments, such as the web. • In such a setting, access of the data is not restricted to a small set of users. • And the application programs that process the data may not have been designed specifically for these data; hence the need for them to have access to both the data and their semantics.

13. The Semantic Web • Unlike databases, hypertext data are designed for human consumption. However, these data are not machine processable. • Hence the call for the Semantic Web [Berners-Lee01]. • Machine processable web data has come to mean “…having semantic metadata and ontologies for web content to enable information access, integration, interoperation and consistency…” Katia Sycara ODBASE’03, November 7, 2003

14. The Layered Cake Architecture • From bottom to top… • Unicode, URI; • XML data, XML schema; • RDF data, RDF schema; • Ontology, vocabulary; • Logic; • Proof; • Trust. …but, who uses what, when??

15. Some Concerns • Hard to develop technologies for computationally demanding tasks, e.g., theorem provers, model checkers, deductive databases,… • Scalability?? • Practitioners tend to not use logical specification languages, e.g., Z, Datalog,… • From bottom to top… • Unicode, URI; • XML data, XML schema; • RDF data, RDF schema; • Ontology, vocabulary; • Logic; • Proof; • Trust. We have to blend carefully technologies with methodologies

16. Towards a Novel Theory of Data Semantics • On-going (and very preliminary) work with Alex Borgida and Yuan An. • Basic premise: If we are going to tackle the problem of data semantics -- again -- we better have a new angle at the problem!

17. The Correspondence Continuum • Consider: • A photo of a landscape is a model with the landscape as subject matter; • A photocopy of the photo is a model of a model of the landscape; • A digitization of the photocopy is a model of the model of the model of the landscape….etc. • Meaning is rarely a simple mapping from symbol to object; instead, it often involves a continuum of (semantic) correspondences from symbol to (symbol to)* object [Smith87]

18. Example XMLSchema For UT CS students RelSchema For UT grad CS students RelSchema For UT CS students Subject matter ERSchema For UT students XMLSchema For UT CS-ECE students RelSchema For UT ECE students

19. Correspondence Graphs • The graph associated with each correspondence continuum has a single anchor, its semantic model. • The semantic model is like a formally represented encyclopedia on a given subject matter; it is application-independent, specified in an expressive knowledge representation language (e.g., OWL.) • For example, a semantic model on Napoleon represents concepts such as Battle, Army, General and historic events, such as the battle of Waterloo. • Every model has an associated (semantic) correspondence to one or more other models. • No cycles are allowed.

20. Models • A model is intended to answer a specific set of questions about its subject matter (a model has a purpose!) [Ladkin97]. • For example, a model airplane can answer questions about the dimensions and aerodynamics of an aircraft; but not questions about its engine power, physical makeup, etc. • For every model, we need a translation function that will translate a query about the subject matter into one about the model (where applicable), and vice versa for the result of the query.

21. Types of Models • I-models (intentional): Consist of a set of predicates with associated axioms. Database schemas, but also logical theories fit here. • E-models (extensional): These have set-theoretic constructions, and query answering based on set-theoretic relationships; Tarskian and Kripke models, but also databases, fit here. • C-models (computational): These are characterized by the fact that query answering is produced by running programs.

22. Correspondences • A correspondence defines the semantics of a model with respect to its subject matter. • This may be done in terms of GAV, (LAV?) or GLAV mappings, maybe others as well. • Correspondences have types too: denotations, representations, implementations, specifications,... • Correspondence composition can be done on the basis of their types.

23. Compositions • The semantics of a model m consists of a composition of the correspondences c1, c2, …, cn that link it to its semantic model. • The whole theory rests on the premise that we can come up with a rich enough class of correspondences for which composition is meaningful and computationally tractable.

24. So, What Does All This Mean? “ontologies” schemas More semantics

25. Remarks • Most computations involve “simple” models, schemas, databases and XML data; some involve their semantics as well. • Mappings and mapping compositions are required here. • Most contributors to the Semantic Web vision get to use well-known database technologies (schemas, queries, views,…). • Semantic web applications resort to expensive semantic computations only when they need to. Claim: Mappings are easier to formalize than concepts

26. This is a version of the Semantic Web with an emphasis on mappings and mapping compositions, rather than rich semantic models

27. Semantic Encapsulation • For this to work, we need to assume that Every model comes with a correspondence to another model • We have accepted behavioural encapsulation, why not semantic one? • Of course, there is a price to be paid in requiring that every model (e.g., schema) comes with its semantics… …But there is a far greater price to pay with legacy data

28. Acknowledgements This presentation is based on research conducted in collaboration with Alex Borgida and Yuan An

29. References • [Berners-Lee01] Berners-Lee, T., Hendler, J., Lassila, O., “The Semantic Web: A new form of Web content that is meaningful to computers will unleash a revolution of new possibilities”, Scientific American, May 2001. • [King02] King, R., “The Story of Civilization and the Rubicon of Smart Data”, Proceedings 28th International Conference on Very Large Databases (VLDB’02), Hong Kong, August 2002.